BreedVision

Architecture

BreedVision leverages transfer learning from a pre-trained ResNet50 architecture integrated with YOLOv5 for precise image localization. The final layers of the network were modified and retrained using a softmax activation to match the specific number of target breeds.

Data Pipeline

Training data was entirely custom-sourced. Using the Google Image Search API, thousands of dog images were programmatically scraped, cleaned, and labeled to build a robust training dataset capable of high-accuracy predictions under a strict latency budget.